Simon 1997: what it reports on intrapatient escalation
Reading notes, with every number sourced to a page
Notes on Simon R, Freidlin B, Rubinstein L, Arbuck SG, Collins J, Christian MC. Accelerated titration designs for phase I clinical trials in oncology. J Natl Cancer Inst 1997;89(15):1138–1147. doi:10.1093/jnci/89.15.1138
Read for the design definitions and the simulation results. The model-fitting sections and the worked trial example were not read. Every number below carries the journal page it came from.
Simon’s cohort designs are numbered 1 to 4 and his intrapatient options are lettered A and B. Neither is this project’s designs A (conventional), B (catch-up) and C (scout-and-backfill).
The question these notes answer
The paper’s designs 2, 3 and 4 change two things at once against design 1: they accelerate the cohort structure, and they turn on intrapatient dose escalation. The two effects can still be separated, because the paper reports the mixed designs — design 1 with intrapatient escalation, and designs 2 to 4 without it (p.1143–1144).
The separation comes out one-sided. Intrapatient escalation moves the number of patients treated at subtherapeutic doses and moves nothing else that the paper measures. Everything that shortens the trial comes from the cohort structure, and specifically from the dose step size.
The two escalation options
| Option | Grade 0–1 last course | Grade 2 | Grade 3 or worse |
|---|---|---|---|
| A | stay | stay | de-escalate one level |
| B | escalate | stay | de-escalate one level |
Option A is conventional practice and has no within-patient escalation at all. Option B is the mechanism this project’s catch-up design uses. Grade 2 is implicit in the Appendix definition, which states only the escalate and de-escalate conditions.
Double steps, with and without intrapatient escalation (3A vs 3B)
Design 3 is one patient per dose level with double dose steps during the accelerated stage, reverting to design 1 at the first first-course DLT or the second first-course grade 2 event. 3A and 3B differ only in whether an individual patient’s dose escalates.
| Metric | 3A, no IPDE | 3B, with IPDE | Difference | Source |
|---|---|---|---|---|
| Patients, mean | not reported | 20.7 | stated as “little or no effect” | 3B p.1141; claim p.1143 |
| Patients, median | not reported | 19.3 | — | p.1141 |
| Cohorts, the paper’s time proxy | not reported | Fig. 3 only | stated as “little or no effect” | Fig. 3 p.1143; claim p.1143 |
| Patients whose worst toxicity was grade 0–1 | 6.5 | 3.9 | B treats 2.6 fewer at subtherapeutic doses | 3A p.1144; 3B p.1142 |
| Patients with grade 3 | 5.7 | 6.8 | B costs 1.1 more | 3A p.1143; 3B p.1142 |
| Patients with grade 4 | 3.2 | 4.3 | B costs 1.1 more | 3A p.1143; 3B p.1142 |
The only quantities resolved numerically between A and B are the toxicity-grade counts. Patients and cohorts are reported as a qualitative claim, not a number.
Standard cohorts, with and without intrapatient escalation (1A vs 1B)
The cleaner isolation, because no acceleration is involved. Design 1 is cohorts of three expanding to six, single dose steps throughout, so 1A against 1B is intrapatient escalation on its own.
| Metric | 1A, no IPDE | 1B, with IPDE | Source |
|---|---|---|---|
| Patients | 39.9 mean, 36.7 median | “no effect … compared with 1A” | 1A p.1141; claim p.1143 |
| Cohorts | baseline | “no effect … compared with 1A” | p.1143 |
| Patients whose worst toxicity was grade 0–1 | 23.3 | 19.3 | p.1143 |
Intrapatient escalation on a conventional cohort structure moves 4 patients out of the undertreated group and changes neither the sample size nor the duration.
Where the acceleration benefit comes from
All four rows are the paper’s default pairings: design 1 with option A, designs 2 to 4 with option B (p.1143).
| Design | Early stage | Step size | Patients, mean | Patients, median | Cohorts against design 1 |
|---|---|---|---|---|---|
| 1 | 3–6 per level throughout | single (~40%) | 39.9 | 36.7 | baseline |
| 2 | 1 per level | single (~40%) | 24.4 | 21.8 | no advantage; design 1 slightly fewer |
| 3 | 1 per level | double (~100%) | 20.7 | 19.3 | substantial saving |
| 4 | 1 per level | double (~100%) | 21.2 | 19.1 | substantial saving |
Design 2 removes 15 patients from the trial and saves no time whatsoever. Design 1 needs slightly fewer cohorts than design 2, “because design 2 sometimes overshoots its target and requires more cohorts at de-escalated levels” (p.1142). The saving appears only with double dose steps: designs 3 and 4 “show substantial savings over designs 1 and 2 because of their use of double dose steps during the initial stage of the trials” (p.1142).
Fewer patients is not a shorter trial. The paper says so in as many words and tabulates cohorts separately in order to show it (p.1142).
What the paper does not report
- No figure or table splits option A from option B. Figures 1 to 4 each carry four panels, one per design, in the default pairings only. The entire A-against-B comparison is four sentences of running text on p.1143–1144.
- No numeric value for cohorts anywhere. Fig. 3 (p.1143) is twenty-bin histograms of the average cohort count across the 20 parameter sets, for the four default designs. A number read off it by eye would still not be an A-against-B number.
- No calendar time in weeks or months. Cohorts is the proxy, defined as “time to completion when there is an excess of patients available for entry in the trial” (Fig. 3 caption, p.1143).
- No separate 3A sample size. The claim covering it is “combining designs 2 through 4 with option A also has little or no effect on the number of patients or cohorts required compared with the same design using option B” (p.1143).
What this settles for the specification
Catch-up helps patients, not the clock: supported in direction, unquantified in size. Catch-up escalation helping participants more than it shortens the trial is what this paper reports, and the participant half has a number attached: 3.9 against 6.5 patients at subtherapeutic doses (p.1142, p.1144). The trial-duration half is a qualitative claim in both contrasts. Cite Simon for the direction; the simulation still has to produce the time saving, because no published number exists to take instead.
The conventional comparator is 3A. One patient per dose level, double dose steps, no intrapatient escalation, reverting to design 1 at the first first-course DLT or the second first-course grade 2. That is the design that captures the acceleration without the mechanism under study, and it is what Objection 3 in Section 27 demands be simulated.
Time is bought with step size, not with who escalates. This is the finding that transfers. The dose escalation factor in Section 13 stops being a parameter and becomes the mechanism, and the scout ladder’s ~3.3× steps are the thing to defend.
Option B patients follow the frontier; the scout defines it. Simon’s intrapatient escalation moves patients up into levels the single-patient cohorts have already opened, which is why it adds no time. The catch-up design in Section 6 is exactly that and inherits the null result. The scout in the scout-and-backfill design is not: no cohort opens the levels it climbs into, and it moves on a weekly interval rather than at a cohort review. The null result applies to the catch-up design and not to the scout.
Nobody else appears to report it either
A literature search for a paper that quantifies the calendar-time effect of intrapatient dose escalation did not find one. Everything below comes from search results and abstracts. None of it has been read.
The one paper that could still carry it is Guo et al. 2025 (IP-CRM, Pharmaceutical Statistics, doi:10.1002/pst.2461). It is the only design paper that puts intrapatient escalation inside a comparative simulation framework. Its stated motivation is patient scarcity — pediatric and rare-cancer trials where accrual limits the design — rather than calendar time, so its tables most likely resolve sample size rather than duration.
The literature that does quantify trial duration attacks a different mechanism. The time-to-event family, beginning with TITE-CRM (Cheung and Chappell, Biometrics 2000;56:1177–1182), reports trial duration directly in its simulations. It shortens trials by letting the next patient enrol before the current one has completed the DLT observation window, not by moving any individual patient’s dose.
That split is itself consistent with the null result above. The field treats trial duration as a problem of accrual and observation windows, which is what the frontier argument predicts: within-patient escalation moves patients through territory that has already been cleared, so it cannot make the frontier advance faster.
One consequence for this project. The time-to-event mechanism is orthogonal to the scout, and is a separate candidate lever on \(\Delta T\) that this specification does not currently carry. It belongs in future work rather than in v1.